← Work
Environments2026-08

Sudoku SFT + RL Environment

A verifiable Sudoku environment for supervised fine-tuning and reinforcement learning. Constraints you can score, not vibes.

Sudoku is a small, fully checkable world. Every grid is either valid or it is not. That makes it a clean harness for teaching models to search, backtrack, and respect hard constraints.

The work covers an SFT corpus plus an RL environment: init, step, reward, eval. Same family as FigmaBench: if you cannot score it, you cannot train it.